Image Denoising Network Based on Subband Information Sharing Using Dual-Tree Complex Wavelet

Kui Liu, Yiping Guo, Benyue Su

Abstract: The difficulty in image denoising arises from the need to recover complex texture regions. Previous methods have struggled with this task, producing images that are too smooth and are prone to aliasing and checkerboard patterns in complex regions. To address these issues, we propose an image denoising network based on subband information sharing using dual-tree complex wavelet. This network combines the spatial and transform domains through dual-tree complex wavelet transform (DTCWT) to capture both spatially structured features and time-frequency localized features for enriching the feature space. To strengthen the recovery in hard scenes, such as weak textures and high-frequency details, Subband Information Sharing Unit (SISU) is designed for the interplay of information in the transform domain, establishing the complementarity and correlation among the subbands obtained by DTCWT. Moreover, rectified linear units and exponential linear units are used in the spatial and transform domains, respectively, to match the properties of elements in different domains. Comprehensive experiments demonstrate the powerful recovery capability of the network for both textured and smooth regions, as well as the competitive results of the network in non-blind/blind image denoising.

External IDs:dblp:journals/npl/LiuGS23